A Deep Learning Approach for Diabetic Retinopathy and Cataract Detection

Abstract

This study describes a composite deep learning-based model of automated retinal image recognition that simultaneously targets to detect diabetic retinopathy (DR) and cataracts. The proposed approach combines both a convolutional neural network (CNN) to extract robust spatial features and a long short-term memory (LSTM) network to model feature dependencies in sequence, which allows detecting both ocular conditions with high accuracy using the same system. Retinopathy and cataracts are among the major causes of vision loss that are irreversible in cases of delay in the diagnosis of diabetes. Traditional diagnostic methods tend to neglect the combined use of feature and pattern recognition and therefore lead to a low degree of diagnostic reliability. In this paper, extensive image preprocessing practices and data augmentation practices are used to increase the model's robustness and generalization. The CNNLSTM model suggested is trained with the help of TensorFlow and Keras and tested on multiple publicly available retinal image datasets, including over 29,000 labeled images, Messidor-2, DDR, ODIR-5K, and APTOS-2019. In general, the current research presents a valid and effective deep learning system that can be used to detect diabetic retinopathy and cataracts early and consistently with a high degree of accuracy (94.5%), sensitivity (92.3%), specificity (96.8%), and, consequently, the F1-score (92.9) under identical experimental conditions and outperforms CNN-only and RNN-only baseline systems. Generally, the proposed research presents a stable and efficient deep learning framework that can be used to identify diabetic retinopathy and cataracts.